Knowledge Files and RAG in Projects
What problem this solves
Every time an employee starts a new conversation with Claude, Claude has no memory of the previous one and no access to your company's internal documents. This forces people to paste in context repeatedly — copying a contract template, a brand guideline, a pricing sheet — before getting useful output.
Knowledge files solve this. Upload a document to a Project once, and every conversation within that Project has access to it automatically, forever, without the user having to paste anything.
What RAG means in this context
RAG stands for Retrieval-Augmented Generation. The concept: instead of Claude relying only on its training knowledge, it can retrieve relevant chunks from documents you supply and incorporate them into its answer.
In the context of Claude Projects, the implementation is straightforward:
- You upload one or more files to a Project's knowledge store.
- When a user asks a question, Claude searches the uploaded documents for relevant content.
- Claude incorporates what it finds into its response, alongside its general knowledge.
You do not need to configure this. Claude handles the retrieval automatically. You just upload the files.
What you can upload
Supported file types include:
- Text documents —
.txt,.md,.csv - PDFs — including scanned documents with selectable text
- Word documents —
.docx - Spreadsheets —
.xlsx(text content is extracted; Claude cannot execute formulas) - Code files —
.py,.js,.ts, and most programming language files - Presentations —
.pptx(text content extracted)
What Claude cannot do with uploaded files:
- Execute code
- Read password-protected documents
- Process embedded images within a PDF (text surrounding images is read; the image itself is not)
- Access links or external references within a document
Size limits and practical constraints
Anthropic's current limits for Project knowledge files:
| Limit | Value |
|---|---|
| Maximum file size | 30 MB per file |
| Maximum files per Project | Approximately 20 files |
| Total knowledge store per Project | ~200,000 tokens (roughly 150,000 words) |
Practical implication: you cannot upload your entire SharePoint library. You need to be deliberate about what goes in. Upload the documents that are genuinely referenced in most conversations — standard contract templates, policy documents, brand guidelines, pricing sheets — not every file that might theoretically be useful.
If a document is too large, split it. A 300-page contract manual should be split into sections; upload the sections most relevant to the Project's purpose.
What to upload — examples by function
Legal Project:
- NDA templates (yours, and standard ones you frequently receive)
- Contract review checklist
- Key clauses reference sheet
- Jurisdiction-specific compliance requirements
- Deal thresholds and escalation rules
HR / Recruitment Project:
- Job description templates
- Interview question bank
- Competency framework
- Approved benefits information
- Onboarding checklist
Marketing Project:
- Brand voice and tone guidelines
- Messaging frameworks for each product line
- Approved headlines and taglines
- Competitor positioning summary
- Campaign brief template
Finance Project:
- Chart of accounts
- Expense policy
- Month-end commentary template
- Financial reporting definitions
IT / Engineering Project:
- Architecture decision records
- Coding standards document
- Incident response playbook
- Key API documentation
How to upload files to a Project
- Open the Project in claude.ai.
- Click Add content or the paperclip/upload icon in the Project settings panel.
- Select files from your computer.
- Wait for the upload to complete — larger files take longer to process.
Files are processed immediately after upload. Claude can reference them in the next conversation without any further action.
To remove a file: open Project settings → Knowledge → click the three-dot menu next to the file → Delete. Removal is immediate; Claude can no longer reference the deleted file in new conversations.
To update a file: delete the old version and upload the replacement. There is no in-place update; versioning is manual.
What Claude does and does not "know" from uploaded files
Claude reads uploaded files when you ask questions that seem relevant to them. It does not memorise them like a human would. Key behaviours:
It will:
- Find and quote relevant sections when asked
- Apply policies and templates from the documents
- Answer questions based on document content
- Flag when information in a document conflicts with its general knowledge
It will not:
- Spontaneously mention document content without a relevant query
- "Remember" document content if the document has been deleted
- Learn from or update the documents based on conversations
Important distinction: uploaded knowledge files are not used to train Anthropic's models. They exist only within your Project, for your Project's members, for the life of the Project. This is covered in more detail in M63 (Data, Privacy and Security).
Practical scenario: contract review at a law firm
A law firm creates a Legal Review Project with four knowledge files:
standard-nda-template.docx— their preferred NDAreview-checklist.pdf— 47-point contract review checklistjurisdiction-guide.pdf— guidance on 12 key jurisdictionsdeal-thresholds.md— escalation rules by deal size and risk level
Every paralegal and associate on the Project can now ask Claude: "Review this NDA against our standard template and flag deviations" — and Claude will compare against the uploaded template, apply the checklist, check jurisdiction-specific concerns, and flag anything above the deal thresholds. Without the knowledge files, each user would need to paste the template and checklist into every session.
The senior partner estimates this saves 40 minutes per NDA review. Across 80 NDAs per month, that is 53 hours of paralegal time.
Key takeaway
Knowledge files turn a generic AI assistant into a domain-specific tool trained on your actual documents. Start with the three or four documents that are referenced most often in your target workflow. Upload, test, refine. The quality of the knowledge store directly determines the quality of Claude's output for that Project.
📖 Official Documentation See this in practice in Anthropic's live support docs: